Transfer learning-driven attention deep residual network for fine dust prediction: a scalable approach for PM_2.5 monitoring in schools

Air pollution poses a severe threat to children’s health, particularly in school environments, where prolonged exposure to fine particulate matter (PM_2.5) can lead to respiratory diseases and cognitive impairments. Deep learning models have demonstrated potential for predicting environmental PM_2.5 levels; however, their effectiveness is constrained by the availability of location-specific training data, making generalization across schools challenging. This study aims to address this limitation by proposing a novel transfer learning (TL) framework based on a deep residual convolutional neural network (ResNet). By leveraging knowledge from well-monitored source schools and adapting it to target schools with limited data, the proposed approach enhances predictive accuracy and generalization. Results demonstrate that freezing 38 layers, out of 41, in the TL framework achieves optimal performance when moderate amounts of target school data are available. Compared to models trained from scratch, the TL-based model achieved an 82.6% reduction in mean squared error, a 61.5% reduction in mean absolute error, and a 25.7% increase in R² for certain school environments. These findings underscore the potential of transfer learning to develop scalable, cost-effective PM_2.5 monitoring systems, offering a practical solution for real-time management of school air quality.